CCF-BSF:CIF: Small: Coding for Fast Storage Access and In-Memory Computing

第1部分：在移动和云计算需求的推动下，对数据存储的需求呈现出急剧增长的趋势，既包括更高的存储密度，也包括访问性能的大幅提高。由访问挑战驱动的一个相关的令人兴奋的新兴趋势是内存计算，即计算从主处理单元卸载到内存，以减少传输时间和能量。未来快速存储访问和内存计算的挑战无法通过传统的存储体系结构来解决，因为传统的存储体系结构不可避免地要牺牲可靠性和容量来换取延迟。这个瓶颈要求创新的研究贡献，可以同时最大化存储密度，访问性能和计算功能。该项目通过开发原则性的数学基础来解决这一迫在眉睫的挑战，这些基础将支持未来的计算系统，这些系统具有处理新的数据密集型应用所需的质量，重点关注基本的性能界限、算法和实用的信道编码方法。该项目的成果将在现代数据驱动和机器学习应用中得到展示，将推进信息科学中的数学技术，并将直接影响未来的计算机系统架构，以满足日益增长的广泛的社会和科学对计算和快速数据处理的需求。此外，该提案还提供了一些更广泛影响的机制，包括通过首席研究员在加州大学洛杉矶分校领导的现有研究中心参与数据存储和内存行业，课程开发和在加州大学洛杉矶分校工程在线硕士项目中引入新的研究生课程，参与本科生研究人员，以及通过调查式文章和教程传播结果。第2部分：该项目有以下三个互补的研究目标：1)发明新的通道代码，用于可靠和快速的存储访问延迟敏感应用，研究跨越一般存储器和特定方案，特别是电阻存储器。所提出的方案将为充满活力的编码主题提供重要的扩展：具有局部性的编码（代数和基于图的）和约束编码；2)发明新的信道码，其解码直接在内存中进行，以同时满足延迟和可靠性方面的竞争要求。在这里，解码器本身受到计算误差的影响，它们本身表现为数据依赖的意义。该分析将导致对数据相关错误具有鲁棒性的边界和实际代码设计。一个范例将是使用空间耦合设计和使用加窗消息传递解码器解码的代码；3)为稳健的内存计算开发新的基本边界、算法和信道代码，重点是量化现代数据驱动应用中使用的统计推断和其他机器学习算法中计算原语的鲁棒性。其中包括基本的性能限制和新的基于编码的方法，以同时对抗潜行路径和计算噪声。分析将包括（嘈杂的）汉明/欧几里得相似度计算的编码，并在实际机器学习应用的背景下进行评估。该项目的成果也将有助于加州大学洛杉矶分校的课程开发，并将为来自代表性不足群体的本科生研究人员提供新的机会。

Part 1:Driven by the needs of mobile and cloud computing, demand for data storage is exhibiting steep growth, both in the direction of higher storage density as well as a simultaneous ambitious increase in access performance. A related exciting emerging trend driven by access challenges is in-memory computing, whereby computations are offloaded from the main processing units to the memory to reduce transfer time and energy. The challenges of future rapid storage access and in-memory computing cannot be addressed by the conventional storage architectures that inevitably trade off reliability and capacity for latency. This bottleneck calls for innovative research contributions that can simultaneously maximize the storage density, access performance, and computing functionality. This project addresses this imminent challenge by developing principled mathematical foundations that will underpin future computing systems possessing qualities necessary to address new data-intensive applications, focusing on fundamental performance bounds, algorithms, and practical channel coding methods. The results of this project will be demonstrated on modern data-driven and machine learning applications, will advance the repertoire of mathematical techniques in information sciences, and will directly impact future computer system architectures to meet the growing and wide ranging societal and scientific needs for computing and rapid data processing. Additionally, the proposal offers several mechanisms for broader impacts, including engagement with data storage and memory industry through the existing research center that the principal investigator is leading at UCLA, curriculum development and the introduction of new graduate courses in the UCLA on-line master's program in engineering, engagement of undergraduate researchers, and dissemination of the results through survey-style articles and tutorials.Part 2: The project has the following three complementary research goals:1) Invention of new channel codes for reliable and fast memory access for latency sensitive applications, with the study spanning general memories and specific schemes for resistive memories in particular. The proposed schemes will offer non-trivial extensions to vibrant coding subjects: codes with locality (algebraic and graph-based) and constrained coding; 2) Invention of new channel codes for which the decoding is performed directly in memory to enable simultaneously satisfying competing requirements on latency and reliability. Here, the decoder itself is subject to computational errors, themselves manifested in a data dependent sense. The analysis will lead to bounds and practical code designs robust to data-dependent errors. An exemplar will be codes designed using spatial coupling and decoded using windowed message passing decoders;3) Development of novel fundamental bounds, algorithms, and channel codes for robust in-memory computing, with the focus on quantifying the robustness of computing primitives in statistical inference and other machine learning algorithms used in modern data-driven applications. These include fundamental performance limits and new coding-based methods to simultaneously combat sneak paths and computing noise. Analysis will include coding for (noisy) Hamming/Euclidean similarity calculations, evaluated in the context of practical machine learning applications.Results from this project will also contribute to the curriculum development at UCLA and will offer new opportunities for the engagement of undergraduate researchers from underrepresented groups.